Segmentation and characterization of tumors in 18F-FDG PET-CT for outcome prediction in cervical cancer radio-chemotherapy

Cervical cancer is one of the most common cancer to affect women worldwide. Despite the efficiency of radiotherapy treatment, some patients present recurrency. Early unfavorable outcomes prediction could help oncologist to adapt the treatment. Several studies suggest that tumor characteristics visible with FFDG PET imaging before and during the treatment could be used to predict post-treatment recurrency. We present a framework for segmentation and characterization of metabolic tumor activity aimed at exploring the predictive value of pre-treatment and per-treatment F-FDG PET images. Thirty-five patients with locally advanced cervix cancer treated by chemoradiotherapy were considered in our study. For each patient, a coregistered PET/CT scan was acquired before and during the treatment and was segmented and characterized with our semi-automated framework. A segmentation process was applied on the baseline acquisition in order to find the metabolic tumor region (MTR). This MTR was propagated to the follow-up acquisition using a rigid registration step. For every patient, 40 features from the two MTRs were extracted to characterize the tumor changes between the two observation points. We identified explanatory characteristics by exploring the threshold which minimizes the p-value computed from the Kaplan-Meier free-disease survival curves. Seven features were identified as potentially correlated with cancer recurrency (p-value<0.05). Results suggest that our method can compute early meaningful features that are related with tumor recurrence.